| Abstract [eng] |
This master thesis aims to improve Consumer Price Index forecasts. Forecasting accurate price levels is important for governments in formulating monetary policy, such as indexing wages, social benefits, or commercial contracts. While traditionally univariate ARIMA models are used to forecast price indices, we propose transformer models. Harmonised Index of Consumer Prices (HICP) for Lithuania is forecasted using monthly HICP rates from 10 European countries together with unemployment rates, Producer Price Index in Lithuania, two dummy variables, and feature engineered variables for the period 1998-2024. Different types of transformer models are tested against linear models (ARIMA, VAR and ARDL) both in univariate and multivariate settings, with forecasting horizons set to one, three and six months. The results show that linear models are more accurate in short-term (1-3 months) forecasts, whereas transformers are more effective in long-term forecasts. Moreover, multivariate models defeat univariate ones, suggesting a demand for further research on covariates impacting European price indices. In addition, it was discovered that multi-head attention layer, a crucial part in the transformer architecture, could be leveraged as an explanatory tool in multivariate forecasts. |
